Part 2: Abstractions, Design, and Testing

Common architectures

In this chapter we'll introduce neural architectures and AllenNLP abstractions that are commonly used for building your NLP model.

The main modeling operations done on natural language inputs include summarizing sequences, contextualizing sequences (that is, computing contextualized embeddings from sequences), modeling spans within a longer sequence, and modeling similarities between sequences using attention. In the following sections we’ll learn AllenNLP abstractions for these operations. All of these torch.nn.Modules can be used in whatever PyTorch code you want, whether or not you use the rest of what’s in AllenNLP.

1Summarizing sequences

2Contextualizing sequences

3Modeling spans in sequences

4Modeling similarities between sequences

5Other common neural network building blocks

  1. Though if you have a large number of spans, more than a hundred or so, doing this will be slow, and you’ll be better off just using an ArrayField.